HK$31.4 billion. That's the size of Zhipu AI's latest placement — and note, this is a placement, not an IPO.

On July 9, Zhipu AI (02513.HK), Hong Kong's "first LLM stock," announced it would place up to 19.78 million new H-shares via general mandate, with total proceeds expected to hit HK$31.41 billion if fully subscribed. For context: Zhipu's IPO last year raised roughly HK$5 billion. This placement is over 6x its IPO size, setting a record for the largest single tech placement on the Hong Kong Stock Exchange in 2026.

Where's the money coming from? The announcement cites "several independent third-party professional investors, institutional investors, and other investors" — market rumors include national team funds, internet giants, and Middle Eastern sovereign wealth funds, though specific names weren't disclosed. Where's it going? Foundation model R&D, compute infrastructure buildout, commercialization expansion, global ecosystem development.

In plain English: buy GPUs, train models, win customers.

Why Raise This Much, This Fast?

Zhipu isn't the first post-IPO LLM company to rush a follow-on offering, but its scale and urgency set a new record. The reason is simple: large model competition is burning cash faster than anyone expected.

The first cash bonfire is compute. Training LLMs in 2026 is no longer "a few hundred GPUs for a few months." GPT-5.6-class models require tens of thousands of GPUs, with a single training run costing hundreds of millions of dollars. Inference costs are even scarier — once a model deploys to enterprise customers with hundreds of millions of daily calls, inference costs grow linearly, even superlinearly. HK$31.4 billion sounds enormous, but at current GPU prices and electricity costs, that buys maybe tens of thousands of high-end cards and two to three years of runway.

The second cash bonfire is talent. Top LLM researchers now command seven-figure USD annual packages in a global bidding war — if you don't pay up, OpenAI, Anthropic, Google, or DeepSeek will. A few dozen people on a top-tier team means hundreds of millions in annual personnel costs alone.

The third cash bonfire is commercialization subsidies. B2B LLM deployment is currently in a "land grab" phase — you have to be cheaper than competitors, better than competitors, better-serviced than competitors to win clients. That means operating at a loss early on, trading subsidies for market share, and only talking about profitability once customer stickiness sets in.

Second Half of the Arms Race: Who Burns Longest, Not Who Burns Fastest

Zhipu's $4B+ placement signals that China's LLM funding arms race has officially entered the second half.

The first half was about "who can build the best model first" — DeepSeek, Zhipu, Moonshot, Baichuan, 01.AI leapfrogging each other with new models every few months, benchmark scores going back and forth. The second half is about "who survives until commercial profitability" — model capability gaps are narrowing, but commercialization gaps are widening, and capital/compute gaps are widening exponentially.

There's another unspoken reason for Zhipu's massive placement at this juncture: window of opportunity. Right now Hong Kong's AI sector is still hot, liquidity is ample, and investors are still willing to pay for the LLM narrative. Wait six months — if industry consolidation hits, if commercialization disappoints, if market winds shift — and raising this kind of money gets much harder.

For Zhipu, HK$31.4 billion is provisions, ammunition, and a shield. In a winner-takes-most market, you don't necessarily need to be first to profitability — but you do need to be the last one standing. If your balance sheet keeps you alive one year longer than your competitors, you win.

The LLM war ultimately becomes a balance sheet war.

See you tomorrow.